A single AI controller now lets a humanoid robot walk, grab things, and catch itself after a stumble, all without switching programs.
Researchers built HANDOFF, a control architecture for the Unitree G1 humanoid that distills three specialist teacher models (one for motion tracking, one for locomotion, one for fall recovery) into a single 29-degree-of-freedom policy. Rather than feeding the robot a dense stream of joint-by-joint commands, the controller takes a compact 10-dimensional task command, blending the teachers' output based on commanded velocity while a binary flag hands full control to the recovery teacher when a fall looks imminent. On the physical G1, the team reports it matches state-of-the-art velocity tracking and covers the largest robust manipulation workspace among the adapted-interface baselines in their own tests, though the paper does not publish the comparison numbers behind either claim. The same trained controller then ran multi-stage, natural-language-directed loco-manipulation tasks in both simulation and on hardware, with no extra data collection or fine-tuning.
Most humanoid robots today juggle separate controllers for walking, manipulating, and recovering from a stumble, and handing off between them mid-task is a common source of jerky, unreliable behavior. Folding all three into one policy that never swaps brains is a real engineering shortcut, especially for a robot meant to take instructions in plain language instead of pre-programmed routines. But "state-of-the-art" and "largest" here are comparisons against the authors' own baselines, not a head-to-head against every published system, so treat the superlatives as suggestive rather than settled.
Humanoid vendors love to show off a robot walking or a robot grabbing a mug in isolation; the trick nobody has fully cracked is doing both, plus catching itself when it trips, without a human resetting the software in between.